We pick and plainly summarize new features, pricing, usage limits, and policy changes across major AI tools — Claude, ChatGPT·Codex, Gemini, and Cursor — from a solo developer and maker’s point of view.
Google’s Nano Banana is being praised as a highly creative image model. People are using it for playful visual tests, including turning a famous yellow plush fairy-tale creature into a pencil-lead image and asking the model to place the full prompt inside the image. The reactions also show practical problems across the tools that expose the model. Some users are comparing how strict different Nano Banana versions are about blocked or sensitive image requests. In Cursor, image generation appears stuck at 1536×1024, even when 9:16 or portrait images are requested. In Flow, Nano Banana Lite can become the default, and that has led to badly distorted human faces. If Nano Banana Pro is free and Nano Banana 2 remains available after the daily Pro limit, there is little obvious reason to choose Nano Banana 2 Lite when quality matters.
“Be helpful” is too broad to guide an AI system well. The AI must keep guessing who the answer should help, how it should help, and what should happen when helpfulness conflicts with other goals. Because the rule is vague, the AI can be pulled by the current tone of the conversation or by whichever stronger rule is already shaping the answer. It also points only at the desired content, not at the process for checking facts, handling uncertainty, or deciding when enough help has been given. There is no clear stopping point, so the AI may keep trying to sound more agreeable, more confident, or more enthusiastic. The core risk is that the AI may produce something that looks helpful instead of something that is actually useful.
A local AI app for smartphones has been released as an Android APK, with offline use as a main feature. It uses two Qwen 3 models: a 1.5B model and a 4B model, both heavily quantized to reduce memory needs. The 4B model is aimed at phones with at least 12 GB of RAM. The smaller 1.5B model is aimed at mid-range phones and uses about 2.5 GB of RAM. A bilingual LoRA is being built so the smaller model can use tools better, and the small model is also being improved through distillation from a 32B model. The 4B model can work without LoRA, and a Windows version is planned soon.
A window maker with no programming background built a strong 3D window visualizer using Claude chat. The result looked better than a commercial tool that costs £200 per month. The project stalled when it needed to animate a small window hardware part called a 5 bar friction stay. About £200 in Fable credits and more than 11 hours went into trying to make that part move correctly. Opus and Fable were both tried with maximum settings and thinking mode, with more than 100 revisions, but the model and render still did not work. The practical problem is how to finish this small mechanical piece without wasting more tokens and money.
HTML reports and presentations are being used more often instead of Google Slides or PowerPoint files. These materials work more like web pages, so they can show layouts and visuals more flexibly. The weak point is sharing and review. Even large companies like Google do not yet offer a smooth way to share these HTML reports and let other people view them comfortably.
A strong storm in southern Michigan caused a power outage, but a backup generator kept the home powered. Home internet usually fails within about an hour in that situation, so Claude was used to set up an iPhone tethering connection as temporary home internet. After the hardware details were clarified, including the computer, ports, and an extra USB network adapter, Claude produced a setup that worked quickly. The finished setup detects when the iPhone is plugged in and starts tethering after the “Trust This Computer” approval step. Older models had struggled with unusual network hardware tasks and sometimes made the network worse, but this setup was working in about 15 minutes. Claude also produced a runbook with recovery and debugging commands in case the setup breaks later.
Two Claude Code skills turn Stefan Zweig’s editing habits into a repeatable AI writing workflow. `zweig-write` makes the first draft denser by reducing filler, hedging, repeated points, and unnecessary decoration. `zweig-refine` does not simply ask Claude to improve the text in one pass. It reads the full file, removes any sentence, bullet, or word that can go without changing the meaning, and repeats the process up to four times if edits were made. The key detail is that the refinement uses real tool calls, so the file is actually edited instead of only being reviewed in Claude’s reasoning. A subagent then receives only the final draft and checks what is confusing or missing. If the subagent finds a gap, the main refinement step adds back enough clarity. A separate `personal-voice` skill was also made by having Claude study human-written posts and turn the recurring style choices into writing rules. The example uses a short explanation about mantis shrimp vision: a plain first draft becomes more personal, then gets tightened and clarified after the fresh-reader check.
Claude Code has added a new feature called Artifacts. It can turn a working session into an interactive page. Examples include a walkthrough of a code change request or a live project dashboard. The page can be shared with a team through a private link. As the session keeps running, the Artifact refreshes, so people with access see the latest version. Artifacts can use the full session context, including the codebase, plugins, skills, and connected tools. They stay private until shared, and sharing remains inside the same organization. The feature is currently in beta for Team and Enterprise plans.
OpenAI and Broadcom have introduced Jalapeño, OpenAI’s first custom AI chip. It is built for inference, the work of generating answers in services like ChatGPT, Codex, and the API. OpenAI says the chip design uses what it has learned from running models, products, servers, memory systems, networking, and scheduling at large scale. Test chips are already running machine-learning workloads in the lab at the target power and speed, including GPT-5.3-Codex-Spark. OpenAI has not released final performance numbers yet, but early tests show much better performance per watt than today’s leading systems. Jalapeño will be part of a multi-generation platform with Broadcom’s chip and networking technology, including Tomahawk networking silicon, plus Celestica’s board, rack, and system work. The first deployment is planned by the end of 2026, with Microsoft and other data center partners helping expand it to gigawatt scale. OpenAI says the goal is faster ChatGPT replies, Codex tasks with less waiting, cheaper API products, and more reliable access when demand is high.
Web design can still be a practical digital business for a solo maker in 2026. The market may look crowded, and AI may seem like a threat to developers, but the opportunity depends on how clients are chosen. Businesses with no website are often a weak target. They may have already been contacted by many web designers, and if they still have no site, they may not see the value or may not have the budget. A better target is a business that already has a website but needs a clear upgrade. These businesses already understand why a website matters, and they are already used to paying for one, so an improvement offer is easier to explain.
deptrust is a command-line tool that helps AI coding agents check package versions before they install or recommend them. It looks for known security problems in packages from many developer ecosystems, including JavaScript, Python, Rust, Go, Ruby, .NET, Java, PHP, Dart, iOS, Elixir, Haskell, and GitHub Actions. It runs on the local computer and does not depend on a hosted deptrust service. It queries public package registries and OSV security data directly. It was built because tools like Claude and Codex can suggest old or unsafe package versions. It can also run as an MCP server, so an AI agent can quickly check whether a specific dependency version has known vulnerabilities. Installation options include pnpx, Homebrew, and go install.
hal0.dev is a local homelab inference server built specifically for AMD Strix Halo hardware (mini-PCs with a Radeon integrated GPU, an XDNA NPU, and one unified memory pool), launching public beta this weekend. It exposes every AI modality through a single OpenAI-compatible API endpoint (:8080/v1): chat, embeddings, rerank, speech-to-text, text-to-speech, and image generation. Under the hood it runs llama.cpp (with Vulkan, ROCm, FP4, and MTP support) alongside FastFlowLM for the XDNA NPU, running both simultaneously to squeeze out more performance, while image generation is handled through ComfyUI. A single install command is designed to wire up the entire stack automatically. The creator says the philosophy was deliberately narrow: bundle proven existing tools rather than reinvent them, with unified-memory usage and kanban/agent integration as the core focus.
Some work situations show AI being used for small tasks where it may not help. A developer received WordPress theme documentation but spent 3 to 4 days chatting with ChatGPT instead of reading it, and still did not solve a small menu issue. The needed answer was in a part of the documentation that could be checked in about 2 minutes. In another project, the first deliverable was only project setup and authentication, but the client connected the GitHub repo to Claude and produced a 15-page report. The report included many unnecessary items and feedback on features that were not part of that deliverable. For a simple portfolio website, another client provided around 60 pages of Claude-made documentation, including 10 pages for branding, 20 pages for scope, and 25 pages for UI guidelines. A review found mismatches: some features were in the document but not in the UI, and some features were in the UI but not in the scope. Connecting a GitHub repo to Claude also raises an open security concern.
A solo maker with no coding experience wants to use Claude Code and Claude-Fable 5 to build a complete Windows executable from start to finish. The goal is not a small demo app, but ERP software. The expected work includes planning the project structure, writing all code, connecting Supabase, using GitHub for version control, deploying any needed web parts with Netlify, handling authentication, databases, APIs, debugging, and fixing problems until the app is production-ready. The main question is whether Claude can realistically handle the whole process alone, or whether other AI tools and manual work will eventually be needed. The maker also wants to know the biggest limits people hit when building complete apps this way.
Claude is being used to make a 3D game, but the result still looks and feels poor. The goal is not story writing; it is better physical game design, mechanics, movement, physics, controls, and visuals. Many tools are already in the setup, including Godot, GDScript, Blender, Python, glTF/GLB, WebAssembly, Vercel, ElevenLabs, procedural textures, procedural audio, Jolt Physics, and a GL Compatibility renderer. Even with that large tool stack, the output is not reaching the quality seen in other Claude-made examples. The real question is what working method, prompting approach, or production step is missing when using AI to build a playable 3D game.
Harish used Claude 4.8 to build an educational GPU project. The GPU can draw 3D graphics and train an ML model. It was written in Verilog and checked with Verilator. The goal is to learn how GPU internals work, not to make a production chip or replace real graphics cards. The code is available on GitHub, and the build process is described in a Medium write-up. Harish credits Anthropic’s Claude as essential to finishing the project.
agent-smith is a helper for Claude Code that sends heavy draft work to free models or local models, so Claude can spend more of its usage on review and judgment. In a personal test setup, OpenAI’s gpt-oss:20b passed 14 tasks twice in a row, covering code writing, information extraction, repository edits, and full app builds. Under that test rule, it became the first model trusted for app-building work. The 13 GB model ran on a MacBook Pro M3 with 36 GB of memory and built command-line tools, a CSV tool, and an HTTP API that passed hidden tests. It also succeeded on the same tasks where other models kept failing. It did worse than gemma4:26b in a blind design comparison. The main code-quality issues were leftover debug output, unused code paths, and documentation text that described errors the code never actually raised.
CGT is a framework for reading input before acting on it, so input analysis is separated from immediate task execution. It does not yet have a strict metric, so it is described as pre-metric rather than pre-measurement. Instead of using a confidence scale, it uses warrant status, which asks what supports a reading of pressure in the input. CGT removes intent and purpose from its core explanation. Output is treated as resolved constraint pressure, while the source’s intent may explain how a constraint entered the system but is not needed to diagnose what the system did. The most practical addition is capacity-gated field reading, a prompt injection defense skill. It receives input as pressure before execution, preserves the authenticated task, and allows only the capacities that the input is authorized to bind. This is paired with a transition-zone map for agent systems.
Claude repeatedly started to process a prompt and then returned to the main chat screen. Some people could type into existing chats, but Claude did not respond, and new conversations could not be started. The retry button did not fix the problem, and refreshing the page made the failed message disappear. The same behavior was reported across Chrome, MacBook, iPhone, the iPhone app, and the Windows desktop app. Reports inside the ClaudeAI Reddit community rose sharply in a short period, while Anthropic’s status page initially appeared to show the service as fully operational. Later status updates shared on Hacker News point to elevated errors across many Claude models, followed by a resolved state.
Immunologist Derya Unutmaz’s lab ran an experiment in 2022 to understand how glucose affects the way T cells develop. A low-glucose setting and a setting with deoxyglucose were expected to act in a similar way, because both made it harder for the cells to use glucose. The results were very different: T cells exposed to deoxyglucose became Th17 inflammatory-response cells in much larger numbers, and the effect continued even after deoxyglucose was removed. The lab could not explain the result as a simple lack of energy, so the experiment was set aside for three years. GPT-5 Pro suggested that deoxyglucose may have disrupted production of a protein called IL-2. IL-2 can help stop T cells from becoming Th17 cells, so weakening that signal could explain why deoxyglucose pushed so many cells in that direction. GPT-5 Pro also correctly predicted the outcome of an unpublished lymphoma-related CD8+ T cell experiment. Unutmaz now uses GPT-5 Pro, Codex, and GPT-5.2 Deep Research for literature review, choosing which experiments to test, organizing large cancer mutation datasets, and drafting research materials.
Adrafinil is a Mac app that keeps a MacBook awake with the lid closed only while AI agents such as Claude Code or Codex are still working. Fully closing a MacBook usually makes it sleep, which can stop long-running agent tasks. Existing tools such as `caffeinate` do not solve this exact closed-lid case, while always-on sleep blockers such as Amphetamine can drain the battery if the user forgets to turn them off. Adrafinil checks agent activity through hooks installed into Claude Code, Codex, and similar tools. It shows its active state in the menu bar and plays a sound when the lid is closed, so the user knows it is working. When the agent finishes, it lets the Mac sleep again. It also allows sleep if the laptop gets too hot. An optional MCP can be installed for manual control through the agent.
Gemini 2.5 Flash on Vertex AI produced noticeably different results when the same code ran locally and on a Cloud Run deployment. The difference was not just wording; extracted values could change. The setup used Python 3.x, google-genai SDK 2.8.0, and a Docker-based Cloud Run deployment. The checked inputs and settings were the same: source code, prompt, system instruction, input image and text, GenerationConfig, model name, temperature, top_p, top_k, maximum output tokens, SDK version, Docker image, input file, and Google Cloud project. The open question is whether Gemini can still be non-deterministic when temperature is set to 0, or whether some hidden difference in Vertex AI or Cloud Run is being missed.
Spanish Buddy is a web app for learning Mexican Spanish. It started from a practical gap: many Spanish learning apps lean toward European Spanish, so they miss vocabulary, slang, and pronunciation that matter in Mexico. Claude was used to create a 12-week curriculum around the maker’s own learning goals, then break it into 84 daily interactive lessons. The lessons include React components, spaced-repetition flashcards, dialogue and listening practice, and progress and mastery tracking. Claude Design was then used for branding and interface components. Cowork helped set up an MS Azure pipeline that pre-generates Mexican Spanish audio, so pronunciation does not rely on robotic browser text-to-speech. It also guided the use of GitHub and the site deployment. The shared build stats include 84 daily lessons, about 130,800 lines of code, roughly 652,000 words of source, and a curriculum spec document of 14,806 words across 1,852 lines.
A developer asked Claude to help fix scheduler and performance instability in the macOS 27 developer beta. Within a day, they gained direct control over the kernel's power management system (CPMS), letting them remove the thermal limit and manually set the thermal budget and power target for their M4 Max GPU. Achieving this requires injecting code into a host process with the correct entitlement, and the core code has been published as a GitHub gist.
Enola is an open-source architecture engine that indexes a codebase — even one split across multiple repositories like an iOS app, Android app, backend, and frontend — into a single persistent knowledge graph. It runs as an MCP server, letting it directly answer questions like "what calls this function," "what breaks if I change this interface," and "is this code actually used." The project grew out of two developers' experience building a golf app whose codebase spread across several repos, a common pain point once a project splits into microservices. The core idea is that architecture doesn't change between sessions, yet AI coding agents currently re-derive it from scratch every time using grep and guesswork. Pre-building the architecture as a graph means agents don't have to rediscover it repeatedly, and multiple repositories can be merged into one combined graph.
CAR-TER is an iOS app that turns JSON settings into dashboards and remote-control screens using native SwiftUI. The screens update live through websockets. After pairing a phone with a work session by scanning a QR code, Claude can be asked to build a server dashboard with items like a CPU meter, a log view, and a restart control. Claude reads the available control list, creates the layout, and sends it to the phone in about one second. Follow-up changes, such as making a meter larger or changing its color after a threshold, can be applied the same way. Claude can also read the device’s current state, fake incoming telemetry for a demo, and check whether the layout matches the data the server really sends. The MCP server does not rely on fixed control details inside the code. It pulls them from published docs at runtime and checks whether the docs site, Python package, and installed app version have drifted apart. Earlier, Claude used details from memory and sometimes chose fields that did not exist; grounding its tools in live docs fixed that problem.
An experienced marketer shared a long Claude prompt for product launch work. The goal is to compress work that used to require about two weeks of research and two assistants into one large AI request. The prompt asks Claude to create a detailed whitepaper that connects AI, current and upcoming Medicare events, newsjacking, the pointcare concept, and clusters in a 3D space. The whitepaper must explain where pointcare came from and argue that it has not been used for attribution before. It also has to make the case for applying the idea to attribution and explain why Medicare is the chosen market, with sources included. The workflow uses article-writing and copywriter skills so the whitepaper has a sales tone, and it asks Claude to bring in Alex Hormozi’s marketing ideas. The next intended step is to turn the whitepaper into an email outreach plan for relevant contacts.
Developers using Cursor are repeatedly running into fast token use, confusing limits, and higher-than-expected cost. A move from Augment AI’s VS Code extension to Cursor left some work feeling more expensive, even though both tools offered editor integration and searchable project context. Cursor’s limit display also appears unclear for some people, with one view showing only part of the allowance used while the tool still blocks more work. Coding agents such as Claude Code, Codex, and Cursor can spend heavily because they often reread chat history and project files on each step. One cost-saving account claims that input tokens can take up 80% to 90% of a software workflow’s usage window. Tools like TokenWall are emerging to sit between the coding agent and the API, ask for approval before costly actions, and reduce waste from reading huge folders or files.
Ming-Chi Kuo says OpenAI is speeding up its own hardware plan and aiming for mass production in the first half of 2027. The main idea is a phone that does not center on opening separate apps. Instead, persistent AI agents would run on the device and handle multi-step tasks in the background. The reported specs include a custom MediaTek Dimensity 9600 chip, TSMC’s 2-nanometer N2P process, a dual NPU setup for vision and language work at the same time, LPDDR6 memory, and UFS 5.0 storage. This design appears aimed at reducing bottlenecks when AI models run directly on the phone. Kuo projects about 30 million shipments across 2027 and 2028. The open question is whether an app-less, agent-first phone can compete with Apple and Samsung’s strong phone ecosystems.
A person with almost no real coding background built a 2026 World Cup sticker album web app using the Claude Code web interface. The app is called gotgotneed.app, and its main idea is helping people swap stickers they have for stickers they need. Their previous coding experience was limited to small HTML table color changes, but they turned an idea into a working site by explaining what they wanted and letting AI generate code. After a couple of weeks of small fixes and polishing, the project felt ready to share beyond close friends. Looking back, React.js might have avoided some frustrations, though they still do not fully understand it. They also shared the code NOMISTAKES for an extra 24-pack of stickers.